The Law of Parsimony: How Occam's Razor Cuts Through Complexity in Science, Crime, and Data

The Principle That Refuses to Go Away

William of Ockham was a 14th-century English friar, a logician, and, by most accounts, a man who had little patience for unnecessary complexity. The philosophical heuristic that bears his name — lex parsimoniae, or the law of parsimony — states that among competing hypotheses, the one requiring the fewest assumptions should be preferred. Seven centuries later, his principle sits at the foundation of machine learning, forensic reasoning, judicial decision-making, and evolutionary biology.

That longevity is not an accident. It reflects something genuinely useful about how we extract reliable conclusions from incomplete data. But the principle is also routinely misunderstood, misapplied, and occasionally weaponized to shut down legitimate complexity. The evidence behind Occam's razor is more nuanced — and more fascinating — than the pop-science shorthand of 'keep it simple' suggests.

What Occam's Razor Actually Means

The common paraphrase — 'the simplest explanation is correct' — is not quite right. The principle does not assert that the world is simple. It asserts that, given two models that explain the same data equally well, the one with fewer free parameters, fewer assumptions, and fewer unverifiable entities is the more rational starting point.

This distinction matters enormously in practice. A complex model is not wrong because it is complex. It is suspect because complexity carries a cost: every additional variable or assumption is another opportunity to be wrong. In scientific practice, parsimonious models are more interpretable, more portable across contexts, and more usable in high-stakes, real-world decision-making than their overly elaborate counterparts.

Some scholars go further, arguing that Occam's razor is the essential demarcation line separating scientific inquiry from superstition and pseudoscience. When an explanation requires invisible forces, untestable entities, or an escalating series of special conditions to account for new evidence, parsimony provides the methodological scalpel to excise it.

The Bayesian Foundation: Why Math Agrees with Ockham

For centuries, the law of parsimony rested on philosophical intuition. In the modern era, it has been given a rigorous mathematical foundation through Bayesian inference — a framework for updating beliefs in light of new evidence.

Bayesian Occam's Razor (BOR) demonstrates that Bayesian inference naturally penalizes complex models. Here is the mechanism: a model with more free parameters can technically fit a wider range of possible data sets. But this flexibility is itself a liability. Because a complex model spreads its predictive probability across a larger space of outcomes, it assigns lower prior probability to any specific observed data set. A simpler model, forced to concentrate its predictions on a narrower range of outcomes, is more strongly confirmed when those outcomes actually occur.

In other words, the mathematics of rational belief-updating independently arrives at the same conclusion Ockham reached philosophically: unnecessary complexity is penalized, not rewarded.

This has direct applications in machine learning. One of the central problems in training predictive models is 'overfitting' — the phenomenon where a model learns the noise in its training data so thoroughly that it fails to generalize to new data. Occam's razor is the conceptual backbone of regularization techniques designed to prevent this. A model that perfectly explains its training set but collapses on new inputs has not learned the signal. It has memorized the static.

The Evidence from the Field

Medical Triage and Decision Trees

Consider the controlled chaos of a mass casualty event. Emergency responders must make rapid decisions about who receives immediate care, who can wait, and who cannot be saved. The tools they rely on are not sophisticated neural networks. They are parsimonious decision trees — simple, interpretable rule sets that can be executed quickly under extreme cognitive load.

Research on medical triage consistently finds that these lean models outperform complex diagnostic algorithms in field conditions. The reason is not that the complex models are less accurate in a controlled laboratory sense. It is that accuracy in deployment requires interpretability, speed, and robustness to missing information. Simplicity, in this context, is not a concession — it is the operationally correct choice.

Recidivism Prediction and the Courts

The criminal justice system has invested heavily in algorithmic risk-assessment tools designed to predict whether a defendant will reoffend. Many of these tools are elaborate, drawing on dozens of variables and complex weighting schemes. The empirical track record is instructive.

Studies in judicial decision-making have repeatedly shown that complex recidivism prediction models often fail to outperform simple, parsimonious models that rely on only a few key variables — typically age at first offense, number of prior convictions, and employment history. The complexity does not add predictive power. It adds opacity. And in a legal context, opacity has consequences: it makes model outputs harder to challenge, harder to explain to juries and judges, and harder to audit for embedded bias.

Evolutionary Biology and Maximum Parsimony

In evolutionary biology, the method of maximum parsimony is a standard tool for reconstructing phylogenetic trees — the branching diagrams that trace the evolutionary relationships among species. The method works by seeking the evolutionary pathway that requires the fewest genetic changes to account for the observed distribution of traits across species.

The logic is explicit: evolution can, and does, produce convergent traits through independent pathways, but invoking multiple independent evolutionary events to explain a shared trait is a stronger claim than invoking common ancestry. Parsimony does not assume evolution is simple. It assumes that the most cautious explanation — the one that makes the fewest unverified claims about independent events — is the right starting point.

Astronomy Before the Telescope

Long before empirical confirmation was possible, parsimony shaped cosmological reasoning. The 14th-century philosopher Jean Buridan argued for a rotating Earth not because he had direct observational evidence, but because Earth's rotation provided a simpler mechanical model for the apparent motion of celestial bodies than the alternative: a stationary Earth requiring every star and planet to independently complete vast orbits in 24 hours. The simpler model required fewer assumptions about the mechanics of the heavens. Centuries later, observation confirmed what parsimony had already recommended.

The Limits of the Razor

Occam's razor is not a law of nature. It cannot be proven from first principles, and it carries no universal guarantee of truth. Nature is under no obligation to be simple — and in many domains, it demonstrably is not. Quantum mechanics, the standard model of particle physics, and the epigenetic regulation of gene expression are all genuine features of the world that are irreducibly complex.

The principle also has a subjective dimension that its advocates sometimes understate. 'Simplicity' is not an objective property of a model. It depends on which features you are counting, which background assumptions you hold, and what you treat as a single entity versus a compound one. Two researchers applying Occam's razor to the same evidence can arrive at different preferred hypotheses depending on their prior frameworks.

Psychological research adds another complication. Humans have a strong cognitive preference for what researchers call 'root simplicity' — the tendency to favor explanations that identify a single, strong root cause over complex, multi-node causal chains. This preference is useful as a heuristic but dangerous when it becomes a cognitive bias. Not every phenomenon has one root cause. Forcing a single-cause explanation onto a multi-causal system is not parsimony — it is oversimplification, and it produces predictions that fail.

The Razor as Methodology, Not Dogma

The most defensible position on Occam's razor is pragmatic. It is a methodological tool, not a metaphysical claim about the universe. It tells us where to start — with the simplest model consistent with the available evidence — and it tells us to require genuine, measurable predictive improvement before accepting a more complex alternative.

In an era of abundant data and increasingly powerful modeling tools, this discipline is more necessary than ever. The capacity to fit a complex model to data has never been greater. So has the capacity to mistake noise for signal, to overfit the past and mispredict the future, to construct elaborate explanations for patterns that do not exist.

Ockham's 14th-century friar could not have anticipated neural networks, Bayesian statistics, or mass casualty triage protocols. But the principle he articulated continues to do real epistemic work in all of them. The evidence, across seven centuries and dozens of domains, suggests he was onto something durable: when in doubt, cut.